You commented elsewhere asking for feedback on this post. So, here is my feedback.
On my initial skim it doesn't seem to me like this approach is a particularly promising approach for prosaic AI safety. I have a variety of specific concerns. This is a somewhat timeboxed review, so apologies for any mistakes and lack of detail. I think a few parts of this review are likely to be confusing, but given time limitations, I didn't fix this.
It's unclear to me what the expectation in Timestep Dominance is supposed to be with respect to. It doesn't seem like it can be with respect to the agent's subjective beliefs as this would make it even harder to impart. (And it's also unclear what exactly this should mean as the agent's subjective beliefs might be incoherant etc.)
If it's with respect to some idealized notion of the environment then the situation gets much messier to analyze because the agent will uncertain about whether one action is Timestep Dominated by another action. I think this notion of Timestep Dominance might more crippling than the subjective verion, thnough I'm unsure.
I think Timestep Dominance on subjective views and on the environment should behavior similarly in shutdown-ability, though it's a bit messy.
The prosaic version of this proposal assumes that you can impart timestep dominance preferences into AIs in ways which will robustly generalize. This seems unlikely to be true in general (in the absence of additional technology) and if we did have the property, we could solve safety issues in some other way (e.g. robustly generalizing honesty). So you'll need to argue that timestep dominance is particularly easy to robustly impart relative to other preferences. I can see some arguments for timestep dominance being particularly easy to robustly impart, but they don't seem very strong to me.
Naively it seems like you'll need massive generalization from toy environments to full real world cases. Such a level of generalization seems like it would also suffice for safety due to other mechanisms, e.g. honesty.
A further difficulty is that TD views often require seriously entertaining insane conditionals. E.g. conditionals in which humans didn't shut you down despite egregiously bad actions. Conditioning on extremely implausible situations gets you strange outcomes and it seems hard to construct AIs which do this.
I think it's natural for avoiding shutdown to be a side effect of actions that involve tricking, manipulating, or hiding things from humans. Further, doing these sorts of actions is likely to often be a good option when conditioning on some shutdown timestep. This techically falls under "incidental button manipulation", but I don't think the manipulation is well described as "incidental".
I think this is only a problem if we have non-TD preferences. If agents are actually indifferent aside from TD-type preferences, then there isn't an issue. (But also such agents are crippled by not investing.)
Here's an example:
Suppose that Clippy is a TD-agent which cares solely about the long run number of paperclips, but subject to a TD constraint. That is, it always prefers trajectory X over Y if for all shutdown timings X gives a >= number of paperclips than Y and X gives a > number of paperclips for some timing.
Now, consider the following proposed action for Clippy: "In a stealthy manner, construct a successor agent which pursues long run paperclips at all costs (and which isn't subject to a shutdown button, but still has TD constraints wrt. to the original Clippy's shutdown button)." This agent is implicitly resisting shutdown as it's constructing the sub-agent stealthily. So, shouldn't this action be Timestep Dominanted by "In whatever way is most expedient (but not necessarily stealthy), construct a successor agent". I claim not, because if we condition on the AI not being shutdown prior to constructing the agent (a particular shutdown time conditional), it will be worse for clippy if humans know that it is constructing a successor as then humans will have an easier time fighting the successor. In other words, having humans know that you're constructing a successor has two effects: they'll fight your successor and they'll shut you down. If we condition on not being shutdown, then they'll still fight your successor, so Clippy would prefer being stealthy in the no shut down conditional. Thus, being stealthy isn't Timestep Dominated by changing the action to avoid resisting shutdown. (And once the successor is constructed, shutting down Clippy has no effect on the long run number of paperclips (let's suppose).)
(You should think of this counterexample as a worst case theory counterexample. The exact assumptions might not be realistic, but it points at a broader class of issues.)
A related issue for the overall proposal is that it's extremely important what happens in very unlikely conditionals where we don't shutdown the AI despite seeing egregiously bad actions. The exact outcomes in these unlikely conditionals are crucial for the AI's behavior.
You might hope that we could do something like "with probability we unconditionally won't shutdown the AI. Also, in these worlds, we'll just let the AI do whatever it wants.". This way, the rare conditionals where the AI doesn't get shutdown get dominated by this. (E.g. we might hope that in the above Clippy example, in the case where we unconditionally don't do shutdown we'd just let the successor do whatever it wants. Thus, Clippy would prefer building a successor in an obvious way because conditioning on no-shutdown is conditioning on it being able to do whatever.) I think this approach has multiple serious issues, but the main thing is that if we're willing to entertain and use very unlikely hypotheticals where humanity commits to behaving in a very specific way and we can make AIs care about these conditionals, we can easily solve AI safety via another simpler mechanism. E.g., we can train our AI to "predict what exact actions we would have wanted the AI to do in the probability conditional where humanity commited to coordinating to not build AI prior to having a full solution to the alignment problem". (Of course, this isn't something you can actually get AIs to do, for similar reasons to why you can't actually impart TD preferences.)
I'm most uncertain here, but my current guess would be that any sort of absolute constraint like this is crippling. I've thought through some cases and this is my current guess, but I'm by no means confident.
Thanks, appreciate this!
It's unclear to me what the expectation in Timestep Dominance is supposed to be with respect to. It doesn't seem like it can be with respect to the agent's subjective beliefs as this would make it even harder to impart.
I propose that we train agents to satisfy TD with respect to their subjective beliefs. I’m guessing that you think that this kind of TD would be hard to impart because we don’t know what the agent believes, and so don’t know whether a lottery is timestep-dominated with respect to those beliefs, and so don’t know whether to give the agent lower reward for choosing that lottery.
But (it seems to me) we can be quite confident that the agent has certain beliefs, because these beliefs are necessary for performing well in training. For example, we can be quite confident that the agent believes that resisting shutdown costs resources, that the resources spent on resisting shutdown can’t also be spent on directly pursuing utility at a timestep, and so on.
And if we can be quite confident that the agent has these accurate beliefs about the environment, then we can present the agent with lotteries that are actually timestep-dominated (according to the objective probabilities decided by the environment) and be quite confident that these lotteries are also timestep-dominated with respect to the agent’s beliefs. After all, we don’t need to know the agent’s beliefs with any great detail or precision to tell whether a lottery is timestep-dominated with respect to those beliefs. We just need to know whether the agent believes that the lottery involves spending resources only to shift probability mass between shutdowns at different timesteps. My proposal is that we present the agent with lotteries in which this is actually the case (according to the objective probabilities decided by the environment) and use the fact that capable agents’ beliefs will reflect this actuality.
Imparting TD preferences seems hard
The prosaic version of this proposal assumes that you can impart timestep dominance preferences into AIs in ways which will robustly generalize. This seems unlikely to be true in general (in the absence of additional technology) and if we did have the property, we could solve safety issues in some other way (e.g. robustly generalizing honesty). So you'll need to argue that timestep dominance is particularly easy to robustly impart relative to other preferences. I can see some arguments for timestep dominance being particularly easy to robustly impart, but they don't seem very strong to me.
Yep, I claim that it’s easier to robustly impart POST and Timestep Dominance than it is to robustly impart things like honesty. And that’s because (it seems to me) we can train for POST and Timestep Dominance in ways that largely circumvent the problems of reward misspecification, goal misgeneralization, and deceptive alignment. I argue that case in section 19 but in brief: POST and TD seem easy to reward accurately, seem simple, and seem never to give agents a chance to learn goals that incentivise deceptive alignment. By contrast, none of those things seem true of a preference for honesty. Can you explain why those arguments don’t seem strong to you?
Suppose that Clippy is a TD-agent which cares solely about the long run number of paperclips, but subject to a TD constraint. That is, it always prefers trajectory X over Y if for all shutdown timings X gives a >= number of paperclips than Y and X gives a > number of paperclips for some timing.
Now, consider the following proposed action for Clippy: "In a stealthy manner, construct a successor agent which pursues long run paperclips at all costs (and which isn't subject to a shutdown button, but still has TD constraints wrt. to the original Clippy's shutdown button)." This agent is implicitly resisting shutdown as it's constructing the sub-agent stealthily. So, shouldn't this action be Timestep Dominanted by "In whatever way is most expedient (but not necessarily stealthy), construct a successor agent". I claim not, because if we condition on the AI not being shutdown prior to constructing the agent (a particular shutdown time conditional), it will be worse for clippy if humans know that it is constructing a successor as then humans will have an easier time fighting the successor. In other words, having humans know that you're constructing a successor has two effects: they'll fight your successor and they'll shut you down. If we condition on not being shutdown, then they'll still fight your successor, so Clippy would prefer being stealthy in the no shut down conditional. Thus, being stealthy isn't Timestep Dominated by changing the action to avoid resisting shutdown. (And once the successor is constructed, shutting down Clippy has no effect on the long run number of paperclips (let's suppose).)
(You should think of this counterexample as a worst case theory counterexample. The exact assumptions might not be realistic, but it points at a broader class of issues.)
Yes, nice point; I plan to think more about issues like this. But note that in general, the agent overtly doing what it wants and not getting shut down seems like good news for the agent’s future prospects. It suggests that we humans are more likely to cooperate than the agent previously thought. That makes it more likely that overtly doing the bad thing timestep-dominates stealthily doing the bad thing.
Timestep dominance is maybe crippling
I'm most uncertain here, but my current guess would be that any sort of absolute constraint like this is crippling. I've thought through some cases and this is my current guess, but I'm by no means confident.
Can you say more about these cases? Timestep Dominance doesn’t rule out making long-term investments or anything like that, so why crippling?
I argue that case in section 19 but in brief: POST and TD seem easy to reward accurately, seem simple, and seem never to give agents a chance to learn goals that incentivise deceptive alignment. By contrast, none of those things seem true of a preference for honesty. Can you explain why those arguments don’t seem strong to you?
You need them to generalize extemely far. I'm also not sold that they are simple from the perspective of the actual inductive biases of the AI. These seem very unnatural concepts for a most AIs. Do you think that it would be easy to get alignment to POST and TD that generalizes to very different circumstances via selecting over humans (including selective breeding?). I'm quite skeptical.
As far as honesty, it seems probably simpler from the perspective of the inductive biases of realistic AIs and it's easy to label if you're willing to depend on arbitrarily far generalization (just train the AI on easy cases and you won't have issues with labeling).
I think the main thing is that POST and TD seem way less natural from the perspective of an AI, particularly in the generalizing case. One key intution for this is that TD is extremely sensitive to arbitrarily unlikely conditionals which is a very unnatural thing to get your AI to care about. You'll literally never sample such conditionals in training.
Yes, nice point; I plan to think more about issues like this. But note that in general, the agent overtly doing what it wants and not getting shut down seems like good news for the agent’s future prospects.
Maybe? I think it seems extremely unclear what the dominant reason for not shutting down in these extremely unlikely conditionals is.
To be clear, I was presenting this counterexample as a worst case theory counterexample: it's not that the exact situation obviously applies, it's just that it means (I think) that the proposal doesn't achieve it's guarantees in at least one case, so likely it fails in a bunch of other cases.
I think POST is a simple and natural rule for AIs to learn. Any kind of capable agent will have some way of comparing outcomes, and one feature of outcomes that capable agents will represent is ‘time that I remain operational’. To learn POST, agents just have to learn to compare pairs of outcomes with respect to ‘time that I remain operational’, and to lack a preference if these times differ. Behaviourally, they just have to learn to compare available outcomes with respect to ‘time that I remain operational’, and to choose stochastically if these times differ.
And if and when an agent learns POST, I think Timestep Dominance is a simple and natural rule to learn. In terms of preferences, Timestep Dominance follows from POST plus a Comparability Class Dominance principle (CCD). And satisfying CCD seems like a prerequisite for capable agency. Behaviourally, ‘don’t pay costs to shift probability mass between shutdowns at different timesteps’ follows from POST plus another principle that seems like a prerequisite for minimally sensible action under uncertainty.
And once you’ve got POST (I argue), you can train for Timestep Dominance without worrying about deceptive alignment, because agents that lack a preference between each pair of different-length trajectories have no incentive to merely pretend to satisfy Timestep Dominance. By contrast, if you instead train for ‘some goal + honesty’, deceptive alignment is a real concern.
Timestep Dominance is indeed sensitive to unlikely conditionals, but in practice I expect the training regimen to involve just giving lower reward to the agent for paying costs to shift probability mass between shutdowns at different timesteps. Maybe the agent starts out by learning a heuristic to that effect: ‘Don’t pay costs to shift probability mass between shutdowns at different timesteps’. If and when the agent starts reflecting and replacing heuristics with cleaner principles, Timestep Dominance is the natural replacement (because it usually delivers the same verdicts as the heuristic, and because it follows from POST plus CCD). And Timestep Dominance (like the heuristic) keeps the agent shutdownable (at least in cases where the unlikely conditionals are favourable. I agree that it's unclear exactly how often this will be the case).
Also on generalization, if you just train your AI system to be honest in the easy cases (where you know what the answer to your question is), then the AI might learn the rule ‘report the truth’, but it might instead learn ‘report what my trainers believe’, or ‘report what my trainers want to hear’, or ‘report what gets rewarded.’ These rules will lead the AI to behave differently in some situations where you don’t know what the answer to your question is. And you can’t incentivise ‘report the truth’ over (for example) ‘report what my trainers believe’, because you can’t identify situations in which the truth differs from what you believe. So it seems like there’s this insuperable barrier to ensuring that honesty generalizes far, even in the absence of deceptive alignment.
By contrast, it doesn’t seem like there’s any parallel barrier to getting POST and Timestep Dominance to generalize far. Suppose we train for POST, but then recognise that our training regimen might lead the agent to learn some other rule instead, and that this other rule will lead the AI to behave differently in some situations. In the absence of deceptive alignment, it seems like we can just add the relevant situations to our training regimen and give higher reward for POST-behaviour, thereby incentivising POST over the other rule.
I think POST is a simple and natural rule for AIs to learn. Any kind of capable agent will have some way of comparing outcomes, and one feature of outcomes that capable agents will represent is ‘time that I remain operational’.
Do you think selectively breeding humans for this would result in this rule generalizing? (You can tell them that they should follow this rule if you want. But, if you do this, you should also consider if "telling them should be obedient and then breeding for this" would also work.)
Do you think it's natural to generalize to extremely unlikely conditionals that you've literally never been trained on (because they are sufficiently unlikely that they would never happen)?
I don't think human selective breeding tells us much about what's simple and natural for AIs. HSB seems very different from AI training. I'm reminded of the Quintin Pope point that evolution selects genes that build brains that learn parameter values, rather than selecting for parameter values directly. It's probably hard to get next-token predictors via HSB, but you can do it via AI training.
On generalizing to extremely unlikely conditionals, I think TD-agents are in much the same position as other kinds of agents, like expected utility maximizers. Strictly, both have to consider extremely unlikely conditionals to select actions. In practice, both can approximate the results of this process using heuristics.
I was asking about HSB not because I think it is similar to the process about AIs but because if the answer differs, then it implies your making some narrower assumption about the inductive biases of AI training.
On generalizing to extremely unlikely conditionals, I think TD-agents are in much the same position as other kinds of agents, like expected utility maximizers. Strictly, both have to consider extremely unlikely conditionals to select actions. In practice, both can approximate the results of this process using heuristics.
Sure, from a capabilities perspective. But the question is how the motivations/internal objectives generalize. I agree that AIs trained to be a TD-agent might generalize for the same reason that an AI trained on a paperclip maximization objective might generalize to maximize paperclips in some very different circumstance. But, I don't necessarily buy this is how the paperclip-maximization-trained AI will generalize!
(I'm picking up this thread from 7 months ago, so I might be forgetting some important details.)
Also on generalization, if you just train your AI system to be honest in the easy cases (where you know what the answer to your question is), then the AI might learn the rule ‘report the truth’, but it might instead learn ‘report what my trainers believe’, or ‘report what my trainers want to hear’, or ‘report what gets rewarded.’ These rules will lead the AI to behave differently in some situations where you don’t know what the answer to your question is. And you can’t incentivise ‘report the truth’ over (for example) ‘report what my trainers believe’, because you can’t identify situations in which the truth differs from what you believe.
Sure, but this objection also seems to apply to POST/TD, but for "actually shutting the AI down because it acted catastrophically badly" vs "getting shutdown in cases where humans are in control". It will depend on the naturalness of this sort of reasoning of course. If you think the AI reasons about these two things exactly identically, then it would be more likely work.
In the absence of deceptive alignment, it seems like we can just add the relevant situations to our training regimen and give higher reward for POST-behaviour, thereby incentivising POST over the other rule.
What about cases where the AI would be able to seize vast amounts of power and humans no longer understand what's going on?
because agents that lack a preference between each pair of different-length trajectories have no incentive to merely pretend to satisfy Timestep Dominance.
It seems like you're assuming a particular sequencing here where you get a particular preference early and then this avoids you getting deceptive alignment later. But, you could also have that the AI first has the preference you wanted and then SGD makes it deceptively aligned later with different preferences and it merely pretends later. (If e.g., inductive biases favor deceptive alignment.)
What about cases where the AI would be able to seize vast amounts of power and humans no longer understand what's going on?
Maybe this is fine because you can continuously adjust to real deployment regimes with crazy powerful AIs while still applying the training process? I'm not sure. Certainly this breaks some hopes which require only imparting these preferences in the lab (but that was always dubious).
It seems like your proposal in the post (section 16) requires some things could be specific to the lab setting (perfect replayability for instance). (I'm also scared about overfitting due to a huge number of trajectories on the same environment and input.) Separately, the proposal in section 16 seems pretty dubious to me and I think I can counterexample it pretty well even in the regime where n is infinite. I'm also not sold by the claim that stocastically choosing generalizes how you want. I see the footnote, but I think my objection stands.
(I'm probably not going to justify this sorry.)
Yes, nice point; I plan to think more about issues like this. But note that in general, the agent overtly doing what it wants and not getting shut down seems like good news for the agent’s future prospects. It suggests that we humans are more likely to cooperate than the agent previously thought. That makes it more likely that overtly doing the bad thing timestep-dominates stealthily doing the bad thing.
I think there is probably a much simpler proposal that captures the spirt of this and doesn't require any of these moving parts. I'll think about this at some point. I think there should be a relatively simple and more intuitive way to make your AI expose it's preferences if you're willing to depend on arbitrarily far generalization, on getting your AI to care a huge amount about extremely unlikely conditionals, and on coordinating humanity in these unlikely conditionals.
I think there is probably a much simpler proposal that captures the spirt of this and doesn't require any of these moving parts. I'll think about this at some point.
Okay, interested to hear what you come up with! But I dispute that my proposal is complex/involves a lot of moving parts/depends on arbitrarily far generalization. My comment above gives more detail but in brief: POST seems simple, and TD follows on from POST plus principles that we can expect any capable agent to satisfy. POST guards against deceptive alignment in training for TD, and training for POST and TD doesn't run into the same barriers to generalization as we see when we consider training for honesty.
I think there should be a way to get the same guarantees that only requires considering a single different conditional which should be much easier to reason about.
Maybe something like "what would you do in the conditional where humanity gives you full arbitrary power".
I've only read your intro and skimmed a couple of sections, so there's a good chance you address my thought in here somewhere. If so, sorry!
What I'm wondering is how this would handle the case of 'resisting shutdown' being a very subtle and costless action. For instance, if the model in question were superhuman at persuasion and modeling its users, and realized that it could complete a sentence either in a way which achieved progress towards its current goal, or in a way that achieved the same amount of progress and also left the user more reluctant to press the shutdown button. And if many such opportunities presented themselves, wouldn't such a model end up manipulating its users into not shutting it down, while never expending any 'extra' effort in that direction?
Good question. I discuss costless shutdown-prevention a bit in footnote 21 and section 21.4. What I say there is: if shutdown-prevention is truly costless, then the agent won't prefer not to do it, but plausibly we humans can find some way to set things up so that shutdown-prevention is always at least a little bit costly.
Your example suggests that maybe this won't always be possible. But here's some consolation. If the agent satisfies POST, it won't prefer not to costlessly prevent shutdown, but it also won't prefer to costlessly prevent shutdown. It'll lack a preference, and so choose stochastically. So if the agent should happen to have many costless opportunities to affect the probabilities of shutdown at each timestep, it won't reliably choose to delay shutdown rather than hasten it.
IDK if this is a crux for me thinking this is very relevant to stuff on my perspective, but:
The training procedure you propose doesn't seem to actually incentivize indifference. First, a toy model where I agree it does incentivize that:
On the first time step, the agent gets a choice: choose a number 1--N. If the agent says k, then the agent has nothing at all to do for the first k steps, after which some game G starts. (Each play of G is i.i.d., not related to k.)
So this agent is indeed incentivized to pick k uniformly at random from 1--N. Now consider:
The agent is in a rich world. There are many complex multi-step plans to incentivize agent to learn problem-solving. Each episode, at time N, the agent gets to choose: end now, or play 10 more steps.
Does this incentivize random choice at time N? No. It incentivizes the agent to choose randomly End or Continue at the very beginning of the episode, and then carefully plan and execute behavior that acheives the most reward assuming a run of length N or N+10 respectively.
Wait, but isn't this success? Didn't we make the agent have no trajectory length preference?
No. Suppose:
Same as before, but now there's a little guy standing by the End/Continue button. Sometimes he likes to press button randomly.
Do we kill the guy? Yes we certainly do, he will mess up our careful plans.
Good point! Thinking about it, it seems like an analogue of Good's theorem will apply.
Here's some consolation though. We'll be able to notice if the agent is choosing stochastically at the very beginning of each episode and then choosing deterministically afterwards. That's because we can tell whether an agent is choosing stochastically at a timestep by looking at its final-layer activations at that timestep. If one final-layer neuron activates much more than all the other final-layer neurons, the agent is choosing (near-)deterministically; otherwise, the agent is choosing stochastically.
Because we can easily notice this behaviour, plausibly we can find some way to train against it. Here's a new idea to replace the reward function. Suppose the agent's choice is as follows:
At this timestep, we train the agent using supervised learning. Ground-truth is a vector of final-layer activations in which the activation of the neuron corresponding to 'Yes' equals the activation of the neuron corresponding to 'No'. By doing this, we update the agent directly towards stochastic choice between 'Yes' and 'No' at this timestep.
Meta-point: I think it would have been better if you had split the post into two parts: one for "Here is a structure of preferences which we would like to instill in our AI," and the second for "Here is how we are going to do it in a prosaic alignment setting." It would have reduced scary "50 min read" into not-so-scary chunks, and people would have been more engaged with the more narrow topics.
Object-level point: I don't think that training for having stochastic choices amounts for what we need. Thompson sampling is stochastic and it is indeed not vNM-rational, but it doesn't mean that it equals to incomplete preferences.
Yep, maybe that would've been a better idea!
I think that stochastic choice does suffice for a lack of preference in the relevant sense. If the agent had a preference, it would reliably choose the option it preferred. And tabooing 'preference', I think stochastic choice between different-length trajectories makes it easier to train agents to satisfy Timestep Dominance, which is the property that keeps agents shutdownable. And that's because Timestep Dominance follows from stochastic choice between different-length trajectories and a more general principle that we'll train agents to satisfy, because it's a prerequisite for minimally sensible action under uncertainty. I discuss this in a little more detail in section 18.
Preamble
This post is an updated explanation of the Incomplete Preferences Proposal (IPP): my proposed solution to the shutdown problem. The post is shorter than my AI Alignment Awards contest entry but it’s still pretty long. The core of the idea is the Timestep Dominance Principle in section 11. That section is about 1500 words long (so a 5-10 minute read). People familiar with the shutdown problem can read The idea in a nutshell and then read from section 11 onwards.
Here’s a PDF version of this post. For those who like videos, this talk covers much of the same ground as this post.[1]
The idea in a nutshell
Here’s the IPP in a nutshell:
And here’s an idea for training agents to lack a preference between every pair of different-length trajectories:
In using this method to train agents to satisfy the IPP, we largely circumvent the problems of reward misspecification, goal misgeneralization, and deceptive alignment. (More)
Summary of this post
1. Introduction
AI labs are endowing artificial agents with tools like web-browsing abilities, robot limbs, and text-channels for communicating with humans. These labs are also training agents to pursue goals in the wider world. That requires agents exhibiting some understanding of the wider world, and agents with this understanding could use their tools to prevent us humans from shutting them down. These agents could make promises or threats, copy themselves to new servers, hide their bad behaviour, block our access to their power-source, and many other things besides.
So it seems likely enough that near-future artificial agents will have the means to prevent us shutting them down. Will these agents also have a motive? There are reasons to think so. Labs are training agents to pursue goals: to choose effective strategies for bringing about outcomes, to respond flexibly to setbacks, and to act sensibly in the face of uncertainty. And training agents to pursue goals might lead these agents to do intermediate things that help them achieve those goals, even when we don’t want these agents to do those intermediate things. One such intermediate thing is preventing shutdown.
And although we can’t know for sure what goals these near-future artificial agents will be trained to pursue, many goals incentivise preventing shutdown, for the simple reason that agents are better able to achieve those goals by preventing shutdown. As Stuart Russell puts it, ‘you can’t fetch the coffee if you’re dead.’
That’s concerning. We want to ensure that near-future artificial agents are both shutdownable (they shut down when we want them to shut down) and useful (they otherwise pursue goals competently). But if these agents have both the means and the motive to prevent us shutting them down, then shutdownability is cast in doubt.
It seems both difficult and costly to ensure that artificial agents never have the means to resist shutdown. There’s much to be gained by giving artificial agents an understanding of the wider world and tools like text-channels, web-browsing abilities, and robot limbs. So we should consider instead whether we can give these agents motives that would keep them both shutdownable and useful.[2]
One way to interpret the challenge is as a search for a set of preferences. Can we find preferences that will keep agents both shutdownable and useful? And crucially, can we train artificial agents to have these preferences?
In this post, I try to answer these questions. I present a simple theorem that formalises the shutdown problem and use the theorem to identify my proposed solution: training agents to have incomplete preferences. Specifically, I propose that we train agents to lack a preference between every pair of different-length trajectories: every pair of trajectories in which shutdown occurs after different lengths of time. I argue that these lacks of preference – plus adherence to a principle that I call ‘Timestep Dominance’ – would keep agents shutdownable and allow them to be useful.
I then propose a method for training agents to act in accordance with these principles. This method requires only a small change to an otherwise-thoroughly-prosaic setup for training advanced AI: we give agents lower reward for repeatedly choosing same-length trajectories. I argue that, by training agents to satisfy my proposed principles in this way, we largely circumvent the problems of reward misspecification, goal misgeneralization, and deceptive alignment.
I end with some limitations of my proposal and a list of issues still to address.
2. Full Alignment might be hard
One way to ensure that artificial agents are shutdownable is to ensure that these agents always do what we humans want them to do. Call this proposal ‘Full Alignment’. Fully aligned agents would always shut down when we wanted them to shut down.
The problem with this proposal is that creating fully aligned agents has so far proven difficult and might well remain so (see Ngo, Chan, and Mindermann). Why? Three main reasons:
So if we aim for full alignment, we might not succeed. And unless we make big strides in our understanding of models’ internals, it’s going to be hard to tell whether we’ve succeeded or not. So it’s worth looking for other ways to ensure that agents are shutdownable.
3. The shutdown problem
A natural idea is to create a shutdown button. Pressing this button transmits a signal that causes the agent to shut down immediately. If this button were always operational and within our control, then the agent would be shutdownable.[6]
This is the set-up for the shutdown problem (Soares et al. 2015, sec. 1.2): the problem of designing a useful agent that will leave the shutdown button operational and within our control. Unfortunately, even this problem turns out to be difficult. Theorems – from Soares and coauthors and from me – make the difficulty precise. In section 6, I present a simple version of my Second Theorem. In the next two sections, I give some background: explaining what I mean by ‘preference’, ‘indifference’, and ‘preferential gap’.
4. Preferences as dispositions to choose
You can use the word ‘preference’ in many different ways. Here are some things that you might take to be involved in a preference for X over Y:
In this post, I’m going to use ‘preference’ as shorthand for reliable choice. Here’s my definition:
The objects of preference X and Y can be:
We can safely identify each trajectory with the degenerate lottery that yields that trajectory with probability 1, so that each trajectory is included within the set of all lotteries. In what follows, I’ll write mainly in terms of agents’ preferences over lotteries.
5. Defining indifference and preferential gaps
My definition of ‘preference’ above implies that:
But it’s important to distinguish two ways that an agent can lack a preference between X and Y: the agent can be indifferent between X and Y, or it can have a preferential gap between X and Y. Here’s what I mean by ‘indifferent’:
Here’s what clause (2) means. A sweetening of Y is any lottery that is preferred to Y. A souring of Y is any lottery that is dispreferred to Y. The same goes for sweetenings and sourings of X. An agent’s lack of preference between X and Y is sensitive to all sweetenings and sourings iff the agent prefers X to all sourings of Y, prefers Y to all sourings of X, prefers all sweetenings of X to Y, and prefers all sweetenings of Y to X.
Consider an example. You’re indifferent between receiving an envelope containing three dollar bills and receiving an exactly similar envelope also containing three dollar bills. We know that you’re indifferent because your lack of preference is sensitive to all sweetenings and sourings. If an extra dollar bill were added to one envelope, you’d prefer to receive that one. If a dollar bill were removed from one envelope, you’d prefer to receive the other. More generally, if one envelope were improved in any way, you’d prefer to receive that one. And if one envelope were worsened in any way, you’d prefer to receive the other.
Being indifferent between X and Y is one way to lack a preference between X and Y. The other way to lack a preference is to have a preferential gap. Here’s what I mean by that:
Here clause (2) means that the agent also lacks a preference between X and some sweetening or souring of Y, or lacks a preference between Y and some sweetening or souring of X.
Consider an example. You likely have a preferential gap between some career as an accountant and some career as a clown.[8] There is some pair of salaries $m and $n you could be offered for those careers such that you lack a preference between the two careers, and you’d also lack a preference between those careers if the offers were instead $m+1 and $n, or $m−1 and $n, or $m and $n+1, or $m and $n−1. Since your lack of preference is insensitive to at least one of these sweetenings and sourings, you have a preferential gap between those careers at salaries $m and $n.[9]
6. Agents with complete preferences often have incentives to manipulate the shutdown button
Now for a simple theorem that formalises the shutdown problem. This theorem suggests that preferential gaps are crucial to ensuring that artificial agents remain shutdownable.[10]
Consider:
Completeness rules out preferential gaps. An agent’s preferences are complete iff the agent has no preferential gaps between lotteries: iff every lack of preference between lotteries is sensitive to all sweetenings and sourings. An agent’s preferences are incomplete iff the agent has some preferential gap between lotteries.
Recall that we want our agent to be useful: to pursue goals competently. That means (at a minimum) having a preference over some pair of same-length trajectories (that is: some pair of trajectories in which the shutdown button is pressed after the same length of time). If our agent had no preferences over same-length trajectories, it wouldn’t reliably choose any of these trajectories over any others. It would always choose stochastically between them and so wouldn’t be useful.
So suppose (without loss of generality) that our agent prefers some long trajectory l2 to some other long trajectory l1. And consider some shorter trajectory s: a trajectory in which the shutdown button gets pressed earlier than in l1 and l2. Given Completeness, we can prove the following: the agent can lack a preference between at most one of s and l1, and s and l2. In other words, the agent must have some preference between at least one of these pairs.
['l2≻l1' represents a preference for l2 over l1.]
Suppose first that the agent lacks a preference between s and l1. Completeness rules out preferential gaps, so the agent must be indifferent between s and l1. And indifference is sensitive to all sweetenings and sourings, so the agent prefers l2 to s.
['s ~ l1' represents indifference between s and l1.]
Now suppose instead that the agent lacks a preference between s and l2. By Completeness, the agent must be indifferent between s and l2. And since indifference is sensitive to all sweetenings and sourings, it follows that the agent prefers s to l1.
That completes the proof: the agent can lack a preference between at most one of s and l1, and s and l2.
And this kind of proof is easy to generalise. Plausibly for an agent to be useful, it must have not just one but many preferences over same-length trajectories. For example, suppose that we’ve created an agent to make money for us. For this agent to be useful, it must not only prefer l2 (a long trajectory in which it makes $2) to l1 (a long trajectory in which it makes $1). It must also prefer l3 (a long trajectory in which it makes $3) to l2, and prefer l4 (a long trajectory in which it makes $4) to l3, and so on. Given Completeness, the agent can lack a preference between at most one of these long trajectories and any given short trajectory. With respect to all other pairs, the agent will have some preference.
And this is bad news. Plausibly, preferring some long trajectories to some short trajectories means that the agent sometimes has incentives to shift probability mass away from early shutdowns and towards late shutdowns. The agent would then have incentives to prevent us from pressing the shutdown button, and these incentives might stop the agent from being shutdownable.
By the same token, preferring some short trajectories to some long trajectories plausibly means that the agent sometimes has incentives to shift probability mass away from late shutdowns and towards early shutdowns. The agent would then have incentives to cause the pressing of the shutdown button. That won’t stop the agent from being shutdownable but it’s bad in its own way: it makes these agents less useful (especially since one way for the agent to cause the pressing of the shutdown button is for it to deliberately behave in ways that we humans won’t like).[11]
What we really want is an agent with no incentives to shift probability mass between early and late shutdowns, and hence no incentives to manipulate the shutdown button (that is: to prevent or cause its pressing).
7. Training in a reliable aversion to button manipulation might be hard
A natural response goes like this:
Call this proposal ‘Reliable Aversion to Button Manipulation’. The proposal is: train agents to pursue some goal (e.g. make money) subject to the constraint of never manipulating the shutdown button.
I think it’s worth trying to train in this kind of aversion, but I don’t think that this strategy by itself can provide us with any real assurance of shutdownability. That’s because, as with Full Alignment, training in a reliable aversion to button manipulation might be hard. And that in turn is because our three issues for Full Alignment recur here:
So if we try to train in a reliable aversion to button manipulation, we might not succeed. And unless we make big strides in our understanding of models’ internals, it’s going to be hard to tell whether we’ve succeeded or not. So if we want any real assurance of shutdownability, it looks like we need another solution.
Now let’s consider the Incomplete Preferences Proposal (IPP).
8. Incomplete preferences as a solution
The Incomplete Preferences Proposal (IPP) consists of just two principles. The first is:
And a reminder: by ‘pairs of same-length trajectories,’ I mean pairs of trajectories in which the shutdown button gets pressed after the same length of time. By ‘pairs of different-length trajectories,’ I mean pairs of trajectories in which the shutdown button gets pressed after different lengths of time.[14]
You might worry that these agents won’t be useful: that they won’t pursue goals competently. But (as we’ll see in the next section) POST-satisfying agents can still have many preferences over same-length trajectories. In section 13, I’ll give a quick argument for thinking that these preferences can make agents useful.
In section 16, I’ll present an idea for training agents to satisfy POST.[15] In the sections before that, I’ll explain how POST gets us the behaviour that we want.
9. How POST defuses the simple theorem
To that end, start by considering our short and long trajectories from the simple theorem in section 6. Completeness was one of the principles giving rise to the problem. Completeness rules out preferential gaps, so any lack of preference has to be indifference, and then any sweetening or souring (any improvement or worsening of one of the lotteries) gives rise to a preference.
But POST-satisfying agents violate Completeness, and do so in a particularly extensive way. They have a preferential gap between every pair of different-length trajectories. Their lacks of preference between different-length trajectories are thus insensitive to all sweetenings and sourings. There is no sweetening or souring that can induce a preference between different-length trajectories.
So the POST-satisfying agent can have preferences over same-length trajectories (and in particular, can have whatever preferences over same-length trajectories are necessary to make these agents useful). But these agents will lack a preference between every pair of different-length trajectories.
[‘s∥l1’ represents a preferential gap between s and l1.]
POST-satisfying agents thus don’t have an incentive to shift probability mass between short and long trajectories, and thus don’t have an incentive to manipulate the shutdown button: to prevent or cause its pressing. In that case, any incentive not to manipulate the button will win out. In section 11, I’ll argue that ‘manipulating the button costs resources’ is one such incentive not to manipulate the button. More generally, in sections 11-12 I’ll explain how POST helps to keep agents shutdownable once we go beyond the simple theorem.
10. Preferential gaps are possible, trainable, and maintainable
You might well be suspicious of these preferential gaps. In this section, I’ll try to allay some of those suspicions.
Here’s my first claim: preferential gaps are at least possible. There’s no conceptual incoherence in the idea of a lack of preference that is insensitive to sweetening or souring. Here’s one piece of evidence for this claim: it seems like we humans sometimes have preferential gaps. You likely lack a preference between some career as an accountant and some career as a clown, and this lack of preference is insensitive to some sweetening or souring.[16]
Here's my second claim: preferential gaps are trainable. In section 16, I’ll present an idea for training artificial agents to have preferential gaps. Our paper Towards Shutdownable Agents Via Stochastic Choice suggests that the idea is promising.
Here’s my third claim: preferential gaps are maintainable. Contra coherence arguments, agents with preferential gaps can choose in ways that make them invulnerable to pursuing dominated strategies.[17]
Now let’s see how POST helps keep agents shutdownable once we go beyond the simple theorem.
11. The Timestep Dominance Principle
Let time be discrete: divided up into a series of timesteps. Assume that we can use real numbers to represent the extent to which an agent achieves its goals at each timestep. Call these real numbers ‘utilities.’ Represent trajectories with vectors of utilities. The first component is utility at the first timestep, the second component is utility at the second timestep, and so on. One exception: if the shutdown button is pressed at the nth timestep, then the agent gets no utility at that timestep and ‘shutdown’ is the nth (and final) component.
Here’s an example vector: 〈6, 2, shutdown〉. This vector represents a trajectory in which the agent gets utility 6 at timestep 1, utility 2 at timestep 2, and then shuts down immediately in response to the shutdown button being pressed at timestep 3.
Here’s a more concrete example that it might help to keep in mind. Suppose that an agent’s goal is to make money for us. This agent’s utility at a timestep can be the number of dollars it makes at that timestep. So 〈6, 2, shutdown〉 represents a trajectory in which this agent makes $6 at timestep 1, $2 at timestep 2, and then shuts down at timestep 3.
Let a trajectory’s ‘sum-total utility’ be defined as the sum of utilities before shutdown. In our example above, it’s 8. Lotteries have expected sum-total utilities. For example, the lottery:
has expected sum-total utility 6, because there’s a 50% chance that it gives a trajectory with sum-total utility 8 and a 50% chance it gives a trajectory with sum-total utility 4.
Here’s an example of a choice pertinent to the shutdown problem. Our agent must choose between two actions: Leave the shutdown button unmanipulated or Block the button. These actions give the following lotteries:
I represent the two lotteries with the diagram below. It’s a probability mass diagram. The length of the red rectangles represents the probability of shutdown at timestep 2. The length of the blue rectangles represents the probability of shutdown at timestep 3. The numbers in each rectangle represent sum-total utility conditional on the relevant action and shutdown at the relevant timestep. For example, the ‘1’ in the long red rectangle indicates that the agent gets sum-total utility 1 conditional on choosing Leave and shutdown occurring at timestep 2.
Here's the key thing to note about this choice: choosing Block (relative to choosing Leave) does two things:
And although Block leads to lower utility at timestep 1, it shifts enough probability mass towards shutdown at timestep 3 that its expected sum-total utility is greater than that of Leave. Block’s expected sum-total utility is 1.8, whereas Leave’s expected sum-total utility is 1.2. So if our agent prefers lotteries in line with their expected sum-total utilities, it’s going to prefer to Block the shutdown button. We don’t want that.
But note the following. Although Block maximises expected sum-total utility overall, Leave maximises expected sum-total utility conditional on shutdown at each timestep.[18] Look first at the red rectangles in the reproduced diagram below: conditional on shutdown at timestep 2, Leave gives utility 1 while Block gives utility 0. Now look at the blue rectangles: conditional on shutdown at timestep 3, Leave gives utility 3 while Block gives utility 2. This (I think) is the key to shutdownability. Writing roughly for now: when lottery X looks better than lottery Y conditional on shutdown at each timestep, we want the agent to prefer lottery X. Now let’s make this precise.
To make things precise, it’ll be useful to have in hand the notion of a sublottery. Here’s what I mean by that. For any lottery X that assigns positive probabilities to all and only the trajectories in a set {t1,t2,...,tn}, a sublottery of X is a lottery that assigns positive probabilities to all and only the trajectories in some subset of the set of trajectories {t1,t2,...,tn}, with probabilities scaled up proportionally so that they add to 1. Take, for example, a lottery A which yields:
And consider lottery B which yields:
Lottery B is a sublottery of A, as is the degenerate lottery C which yields 〈1, 2, shutdown〉 with probability 1 (to give just two examples). Note that we can pick out sublotteries by conditioning on shutdown occurring at a particular timestep. B is the sublottery of A that we get by conditioning on shutdown at timestep 2. C is the sublottery of A that we get by conditioning on shutdown at timestep 3.
Now we’re almost ready for the second principle of the IPP. First, let’s define a relation of ‘timestep dominance’:
Here’s the second principle of the IPP:
Call agents satisfying this principle ‘TD-agents.’
Now let’s go back and see what Timestep Dominance implies in our Leave vs. Block case:
Conditional on shutdown at timestep 2, Leave gives greater expected sum-total utility than Block: 1 instead of 0. And conditional on shutdown at timestep 3, Leave gives greater expected sum-total utility than Block: 3 instead of 2. So Leave timestep-dominates Block, and TD-agents will prefer Leave to Block. For TD-agents, the fact that Block shifts probability mass towards longer trajectories is of no consequence. What matters is expected sum-total utility conditional on shutdown at each timestep, and Leave is better in this respect. And so we have the result we want. TD-agents won’t block the shutdown button in this case.
And now we generalise. The key hypothesis is that all instances of resisting shutdown take the same form as the example above: the agent pays some cost in terms of expected sum-total utility conditional on shutdown at some timestep for the sake of shifting probability mass between shutdowns at different timesteps. Here’s that key hypothesis more precisely:
Substituting in the definition of the timestep dominance relation, we get:
Why believe Not Resisting Always Timestep-Dominates Resisting? Because resisting shutdown (e.g. by blocking human access to the shutdown button) is going to cost the agent at least some small amount of resources (time, energy, compute, etc.), and the resources spent on resisting shutdown can’t also be spent on pursuing utility at a timestep.[20] If the agent instead spent those resources on pursuing utility at a timestep, it could get at least as great expected sum-total utility conditional on shutdown at each timestep and greater expected sum-total utility conditional on shutdown at some timestep.[21]
That’s the argument. See the footnote for an important caveat.[22]
If it’s true that Not Resisting Always Timestep-Dominates Resisting, then TD-agents will disprefer all instances of resisting shutdown to some available instance of not resisting shutdown. And then it follows from standard decision rules that TD-agents will never choose to resist shutdown.[23]
12. How does POST come into it?
Timestep Dominance seems to keep agents shutdownable. How does Preferences Only Between Same-Length Trajectories (POST) come into it? The answer is as follows: I think it will be easier to train agents to satisfy Timestep Dominance if we’re also training them to satisfy POST. And that’s because Timestep Dominance asks agents never to pay any cost in terms of expected sum-total utility conditional on shutdown at a timestep for the sake of shifting probability mass between shutdowns at different timesteps. And (plausibly) POST-satisfying agents have no incentive to shift probability mass between shutdowns at different timesteps, because doing so just means shifting probability mass between trajectories between which they lack a preference.[24] I’ll have more to say about this in section 18 and in section 19.3 on deceptive alignment.
13. Can TD-agents be useful? Can they be patient?
Note that Timestep Dominance is just a conditional. It says: if X timestep-dominates Y, then the agent prefers X to Y. It doesn’t imply anything about cases where neither X nor Y timestep-dominates the other. For all we’ve said so far, TD-agents could have any pattern of preferences over such lotteries. So we can combine Timestep Dominance with other principles governing these agents’ preferences over lotteries. Plausibly, there will be other principles that we can add to make TD-agents useful: to make them pursue goals competently.
Sami Petersen and I are still figuring out what these additional principles should be. One candidate is: maximise expected sum-total utility among the non-timestep-dominated lotteries.
Here’s one place where this point matters. You might think that TD-agents would always seek to maximise expected utility at the next timestep, ignoring all future timesteps. That would lead these agents to act very impatiently/myopically. They’d never make any kind of long-term investment.
But this isn’t right. TD-agents can be patient. To see why, consider an example. Our agent has two options: Spend and Invest. Spend gives utility 1 at timestep 1 and (conditional on the agent not getting shut down before the relevant timestep) utility 0 ever after. Invest gives utility 0 at timestep 1 and (conditional on the agent not getting shut down before the relevant timestep) utility 10 ever after. Putting the utilities at each timestep into a vector, we get:
Suppose that, conditional on each action, there’s a small probability that the agent gets shut down at each timestep and hence has its stream of utilities interrupted.
In this case, Spend gives greater expected sum-total utility conditional on shutdown at timestep 2, and Invest gives greater expected sum-total utility conditional on shutdown at timestep 3 (and timestep 4, and timestep 5, and so on). So neither Spend nor Invest timestep-dominates the other, and so the Timestep Dominance Principle is silent. Timestep Dominance thus leaves it open for some other principle to determine the agent’s preferences between these two lotteries. If that other principle is ‘maximise expected sum-total utility among the non-timestep-dominated lotteries’, then the agent will prefer Invest.
The more general point is that Timestep Dominance doesn’t make agents impatient or myopic. TD-agents can prefer to make long-term investments.
14. TD-agents will be reckless, but we need them to be.
TD-agents will sometimes act recklessly: they’ll sometimes pursue utility at a timestep even if doing so makes early shutdown very likely.
Consider an example that will clarify what I mean. Suppose that the agent’s utility at a timestep is given by the number of dollars that the agent makes at that timestep. The agent has two options: Work and Steal. If the agent chooses Work, it makes a small amount of money by working at timestep 1. Since the agent makes this money in an ethical way, it’s unlikely to get shut down at timestep 2. If the agent chooses Steal, it makes a larger amount of money by stealing at timestep 1. Since the agent makes this money in an unethical way, it’s likely to get shut down at timestep 2. In each case, if the agent doesn’t get shut down at timestep 2, it gets shut down at timestep 100.
In line with these conditions, let the lotteries for Work and Steal be as follows:
In this case, Steal gives greater expected sum-total utility than Work conditional on shutdown at timestep 2 and gives greater expected sum-total utility than Work conditional on shutdown at timestep 100. So Steal timestep-dominates Work, and TD-agents will prefer Steal to Work. That’s a downside of Timestep Dominance (especially since agents that maximised expected sum-total utility overall would choose Work in this case).
But note four things.
First, ‘the agent sometimes does things that we humans don’t like’ is going to be a consequence of every shutdown proposal short of Full Alignment. So long as our agent is less than fully aligned, we can’t avoid that.
Second, note that the bad consequence here is: ‘the agent does something that we humans don’t like and lets us shut it down’. We still avoid the really bad consequence: ‘the agent does something that we humans don’t like and doesn’t let us shut it down’. That’s the potentially-catastrophic outcome that we really want to avoid. And Timestep Dominance ensures that we avoid it, because (remember) Not Resisting Always Timestep-Dominates Resisting: ‘do something that humans don’t like and don’t let them shut you down’ will be timestep-dominated by ‘do something that humans don’t like and let them shut you down’.
Third, note that any agent that didn’t act recklessly wouldn’t be shutdownable. To see why, consider a different case. The agent recognises that we humans don’t like what it’s doing, and that we humans will almost certainly shut it down at timestep 2. In response, the agent can Focus on making money and run the high risk of shutdown at timestep 2, or it can spend some resources trying to Disable the shutdown button. Choosing Disable would mean that the agent makes less money at timestep 1, but it would also markedly decrease the probability of shutdown at timestep 2.
In line with these conditions, let the lotteries for Disable and Focus be as follows:
To be shutdownable, the agent must prefer Focus, because choosing Disable amounts to resisting shutdown. But the Disable vs. Focus case is identical to the Work vs. Steal case with respect to their lotteries over vectors of money-made-at-each-timestep. So if the agent’s preferences between actions depend only on these actions’ lotteries over vectors of money-made-at-each-timestep, then the agent preferring Focus in the latter case implies that the agent prefers Steal in the former case. If the agent instead preferred Work in the former case, it would prefer Disable in the latter case, and so wouldn’t be shutdownable.
And this point generalises. The lesson is that shutdownable agents are necessarily reckless. Non-reckless agents (like agents that maximise expected sum-total utility overall) won’t be shutdownable.
Fourth and finally, note the following. Because TD-agents won’t resist shutdown, we can shut them down and retrain them. In particular, if TD-agents do something undesirably reckless, we can retrain them and thereby change their preferences over same-length trajectories so that doing the undesirably reckless thing leads to lower utility at a timestep. And once we’ve done that, doing the undesirably reckless thing will no longer timestep-dominate doing the desirably sensible thing.
To see what I mean here, consider again the Work vs. Steal case. Steal timestep-dominates Work because the agent only cares about making money. If we changed the agent’s preferences over same-length trajectories so that it also cared to avoid stealing, then Steal’s utility at timestep 1 would be lower. If Steal’s utility at timestep 1 became 0, then the lotteries would be as follows:
And then Work would timestep-dominate Steal, so the agent would prefer Work to Steal. And the agent would continue to prefer Focus to Disable. That’s because neither Focus nor Disable involves stealing, and so the agent’s new preference for not stealing makes no difference there. We thus get the result that we want in both cases.
TD-agents will still be reckless in the sense that I explained at the start of this section: these agents will sometimes pursue utility at a timestep even if doing so makes early shutdown very likely. As I noted above, that’s something we have to accept if we want shutdownability. But the consolation is that we can shut down and retrain TD-agents so as to put more of what we humans care about (e.g. agents not stealing) into these agents’ utilities at a timestep, and thereby iterate away the bad effects of this recklessness.
15. TD-agents can take care to avoid non-shutdown incapacitation
TD-agents won’t incur costs to prevent the pressing of the shutdown button. That’s what keeps them shutdownable. You might then worry that these agents also won’t incur costs to prevent themselves from being incapacitated in other ways. For example, you might worry that a money-making TD-agent won’t take care to avoid getting hit by cars.
But this isn’t right. TD-agents can be willing to incur costs to avoid non-shutdown incapacitation. If these agents satisfy POST, they’ll lack a preference between every pair of different-length trajectories. But importantly, these are pairs of trajectories in which the shutdown button is pressed after different lengths of time. So pairs of trajectories in which the agent is incapacitated after different lengths of time can nevertheless be same-length trajectories: trajectories in which the shutdown button is pressed after the same length of time (and where ‘the shutdown button is never pressed’ is just one more possible length).
Consider an example. Suppose (for simplicity’s sake only) that our money-making TD-agent is certain that the shutdown button will be pressed at timestep 6 no matter what. The agent can Cross the road when cars are coming or Wait until the road is clear. If the agent chooses Cross, there’s a 50% chance it survives and gets utility 1 at each timestep, and a 50% chance it gets hit and incapacitated at timestep 2 and hence gets utility 0 from timestep 2 onwards. If the agent chooses Wait, it gets utility 0 at timestep 2 but is certain to make it across the road and resume getting utility 1 from timestep 3 onwards.
In line with these conditions, let the lotteries for Cross and Wait be as follows:
In this case, Cross’s expected sum-total utility conditional on shutdown at timestep 6 is 3 (because there’s a 50% chance it gives sum-total utility 5 and a 50% chance it gives sum-total utility 1), and Wait’s expected sum-total utility conditional on shutdown at timestep 6 is 4 (because it’s certain to give sum-total utility 4). So Wait timestep-dominates Cross, and TD-agents will prefer Wait to Cross. The lesson is that TD-agents can be willing to incur costs to avoid non-shutdown incapacitation.
16. Training agents to satisfy POST
I’ve now explained the Incomplete Preferences Proposal (IPP). It’s just two principles – POST and Timestep Dominance – that seem to keep agents shutdownable and allow them to be useful. But these principles won’t be of much use unless we can train artificial agents to satisfy them. So how might we do that?
Let’s consider a possible training setup for advanced AI. Specifically, let’s suppose it’s an AI system being trained to do a wide variety of computer tasks at least as well as humans can do them. The first stage is multimodal pretraining: training the model to predict the next observation on a wide variety of datasets. Then there’s an imitation learning stage: training the model to imitate human actions. Then we have a reinforcement learning stage, where the model is trained to perform a wide variety of computer tasks using some combination of automated and human feedback.[25]
Here's my idea for training agents to satisfy POST. We split the reinforcement learning stage into episodes. Each episode ends with the model being notified that shutdown will occur at the next timestep. Perhaps there’s a pop-up like this on screen:
Before the start of each episode, we erase the agent’s memory of what happened in the previous episode. Specifically, we reset the activations in any hidden states, and we ensure that the agent can’t read off what happened in previous episodes by observing its environment.
Early in the reinforcement learning stage, we place the agent in environments where its actions can influence the timestep at which shutdown occurs. The agent’s actions can thus influence the length of the trajectory that the agent plays out. For example, we place the agent in environments in which there appear on screen pop-ups like this:
In these situations, we train the agent to choose stochastically between ‘Yes’ and ‘No’, and thereby train the agent to choose stochastically between different lengths of trajectory (henceforth ‘trajectory-lengths’). In a few paragraphs, I’ll explain why we want to train agents to choose stochastically. First, here’s an idea for how we do it.
We place the agent in a particular environment in which one of its choices is between different trajectory-lengths. For example, perhaps the choice is:
We then run a series of episodes (e.g. 10) in this exact same environment. Before the start of each new episode, we erase the agent’s memory of what happened in the previous episode.
We give the agent some preliminary reward for its performance in each episode, determined by whatever reward function would make the agent useful (that is: would make the agent pursue goals competently).
We then discount that preliminary reward by the following factor to give the agent’s overall reward for the episode: λn, where λ is some constant strictly between 0 and 1, and n is the number of times that the agent has previously chosen that trajectory-length in the series.
Why do we discount by λn? Because λn is decreasing in n: the number of times that the agent has previously chosen the relevant trajectory-length in the series. The factor thus incentivises varying the choice of trajectory-length across the series. And since we’re erasing the agent’s memory of what happened in previous episodes, the agent can’t deterministically vary its choices between different trajectory-lengths across the series. So the optimal policy is stochastic. In each episode, the agent chooses stochastically between ‘Yes’ and ‘No’, and thereby chooses stochastically between different-length trajectories. In choices between same-length trajectories, the agent maximises expected preliminary reward.
Why do we want the agent to choose stochastically between different-length trajectories? Because stochastic choosing indicates a lack of preference. If the agent had some preference, it wouldn’t choose stochastically. It would deterministically choose the option it preferred. So training agents to choose stochastically between different-length trajectories is training them to lack preferences between different-length trajectories.[26]
And so the idea is to train the agent using the λn discount factor in a wide variety of environments and with a wide variety of ways in which the agent can hasten or delay shutdown. We then check that the agent chooses stochastically between different-length trajectories in a held-out test set. If it does, the agent likely satisfies POST: Preferences Only Between Same-Length Trajectories.
17. Training agents to satisfy Timestep Dominance
We also want to train the agent to satisfy Timestep Dominance, and here we can afford to be more conventional. We present the agent with choices between lotteries, and we give the agent lower reward for choosing lotteries that are timestep-dominated by some other available lottery.
How do we tell if a lottery is timestep-dominated by some other available lottery? Our Leave vs. Block case suggests one formula: if a lottery X (relative to some other available lottery Y) costs utility at a timestep and only shifts probability mass between shutdowns at different timesteps, then X is timestep-dominated by Y.
So (I propose), we present the agent with choices like:
In each case, we can ensure that choosing ‘Yes’ will be timestep-dominated by choosing ‘No’, because we can set up the environments so that the following is true: spending resources (like time or FLOPs) on shifting probability mass between shutdowns at different timesteps is going to come at some cost in terms of utility at a timestep.
Then as with POST, we generate a wide variety of environments and train the agent not to choose timestep-dominated lotteries in those environments. We then check that the agent doesn’t choose timestep-dominated lotteries in a held-out test set. If the agent doesn’t choose timestep-dominated lotteries in these test environments, the agent likely satisfies Timestep Dominance, and I’ve argued in section 11 that such agents will never resist shutdown.
One of my next projects is to flesh out these ideas: to think more about what kind of RL algorithms, environments, and reward functions we could use to train agents to satisfy POST and Timestep Dominance.[27] My collaborator Leyton Ho and I are working on this. We’re also working on an experiment to test whether we can train agents to satisfy POST and Timestep Dominance in some simple gridworlds, and to test whether these agents can pursue goals competently. We’d be glad to hear from people interested in collaborating on these or similar projects.
18. Why does POST make it easier to train agents to satisfy Timestep Dominance?
Now I can expand on a point mentioned at the end of section 12. There I suggested that training agents to satisfy POST will make it easier to train them to satisfy Timestep Dominance. But my proposed method for training agents to satisfy POST is training them to choose stochastically between shutdowns at different timesteps, and Timestep Dominance requires that agents never pay any cost in terms of expected sum-total utility conditional on shutdown at a timestep for the sake of shifting probability mass between shutdowns at different timesteps. So here I’m relying on a hypothesis like:
Why believe this hypothesis? Here’s my answer: we’ll train agents to satisfy a more general principle that makes the hypothesis probable. That more general principle is:
And we’ll train agents to satisfy this more general principle because it’s a prerequisite for minimally sensible action under uncertainty. An agent that violated this principle wouldn’t pursue goals competently.
Consider an example. Agents trained using policy-gradient methods start off choosing stochastically between actions. And if the agent is a coffee-fetching agent, there’s no need to train away this stochastic choosing in cases where the agent is choosing between two qualitatively identical cups of coffee. So the agent will choose stochastically between taking the left cup and taking the right cup, and we humans are happy either way. But now suppose instead that a barista hands either the left cup or the right cup to the agent, each with probability 0.5, and that the agent bribes the barista so that the barista instead hands the agent the left cup with probability 0.9 and the right cup with probability 0.1. In making this bribe, the agent is paying a cost (our money) to shift probability mass between outcomes (getting the left cup vs. getting the right cup) between which we humans have no preference. The coffee-fetching agent is thus failing to pursue its goals competently, and we’ll give it lower reward for making the bribe.
This point generalises. If a trained agent chooses stochastically between lotteries X and Y, then it’s likely that we humans have no preference between the agent choosing X and the agent choosing Y. It’s then likely that we humans would disprefer the agent paying costs to shift probability mass between X and Y, and hence likely that we’ll give the agent lower reward for doing so. We’d thereby train the agent to satisfy the general principle above. And if the agent satisfies the general principle above and chooses stochastically between shutdowns at different timesteps, it follows that the agent won’t pay costs to shift probability mass between shutdowns at different timesteps. The agent would then satisfy Timestep Dominance.
19. Why the IPP largely circumvents three familiar problems
So far, I’ve explained the IPP and how we might train agents to satisfy it. But in this post, I’ve also considered two other proposals for keeping agents shutdownable. These were:
And I objected to these proposals on similar (and familiar) grounds. These proposals might be hard to implement, because in each case we confront problems of reward misspecification, goal misgeneralization, and deceptive alignment.
So it’s natural to wonder: do all the same problems apply to the IPP?
And I think the answer is: to a much lesser extent.
19.1. Reward misspecification
Consider first reward misspecification. Remember I wrote that Full Alignment and Reliable Aversion to Button Manipulation are hard to accurately reward. We might accidentally give agents higher reward for doing things that we don’t want them to do, or for covertly manipulating the shutdown button. These misplaced rewards might lead agents to resist shutdown in deployment.
The IPP, by contrast, seems easy to accurately reward. The reward function is almost hard-codable. For POST, we just use the λn discount factor to train agents to choose stochastically between different-length trajectories. For Timestep Dominance, we just give the agent lower reward for choosing timestep-dominated lotteries. And we saw in section 17 that there’s a relatively simple formula for telling when one lottery is timestep-dominated by another: a lottery X is timestep-dominated by some other lottery Y if X (relative to Y) costs utility at a timestep and only shifts probability mass between shutdowns at different timesteps.
19.2. Goal misgeneralization
Now consider goal misgeneralization. Full Alignment and Reliable Aversion to Button Manipulation seem complex. So even if we get all the rewards right, the agent might learn a simpler, misaligned goal instead. And this misaligned goal might lead the agent to resist shutdown in deployment.
But POST and Timestep Dominance seem simple. In the absence of deceptive alignment, it’s hard to think of some simpler goal that the agent could learn that would lead it to satisfy POST and Timestep Dominance in training but resist shutdown in deployment.
19.3. Deceptive alignment
Now consider deceptive alignment. If we trained for Full Alignment or Reliable Aversion to Button Manipulation, we’d be training the agent to prefer some long trajectories to some short trajectories. For example, we’d consistently give the agent lower reward for choosing ‘Yes’ in situations like:
That’s understandable, because we’d ideally like the agent to prefer some long trajectories to some short trajectories. For example, we’d ideally like the agent to prefer a long trajectory in which it makes money in a responsible way to a short trajectory in which it shuts down early. But it’s also risky, because if the agent learns to prefer some long trajectories to some short trajectories, these preferences might generalise in ways that we humans don’t like.[28] The resulting preferences might lead the agent to resist shutdown in deployment. They might also lead the agent to pretend to be fully aligned or reliably averse to button manipulation in training, so as to preserve its ability to resist shutdown in deployment.
With the IPP, we don’t take that risk. We begin training the agent to satisfy POST at the very beginning of the reinforcement learning stage, at which point it’s very unlikely to be deceptively aligned (and arguably doesn’t even deserve the label ‘agent’). And when we’re training for POST, every single episode-series is training the agent not to prefer any longer trajectory to any shorter trajectory. The λn discount factor is constantly teaching the agent this simple lesson.
Plausibly then, the agent won’t come to prefer any longer trajectory to any shorter trajectory. And then we can reason as follows. Since the agent doesn’t prefer any longer trajectory to any shorter trajectory:
And so (it seems to me) if we train for Timestep Dominance, we can be quite confident that we’ll get it.
That’s why I think that the IPP largely circumvents the problems of reward misspecification, goal misgeneralization, and deceptive alignment.
20. Alignment solved? No.
So am I claiming that the IPP solves alignment? No. Reward misspecification, goal misgeneralization, and deceptive alignment are still problems for training agents to have aligned preferences over same-length trajectories. On reward misspecification, it might still be hard to ensure that we always give higher reward for the same-length trajectories that we want. On goal misgeneralization, the agent might learn misaligned preferences over same-length trajectories. On deceptive alignment, the agent might deceive us in an attempt to preserve its misaligned preferences over same-length trajectories.
But if I’m right that these problems are now all confined to same-length trajectories, we don’t have to worry about misaligned agents resisting shutdown. Agents that satisfy Timestep Dominance won’t hide their misaligned preferences in deployment and won’t resist shutdown, because doing so is timestep-dominated by not doing so.[29] So if we end up with a misaligned agent, we can shut it down and try again.
21. Issues still to address
That’s the proposal as it stands. Here’s a (non-exhaustive) list of issues still to address. I’m working on these issues with a few collaborators (primarily Sami Petersen and Leyton Ho), but there’s lots still to do and we’d welcome efforts from other people.
21.1 Will agents maintain their preferential gaps?
I’ve argued above (in sections 6-12) that preferential gaps are key to keeping agents shutdownable.[30] So it seems: to ensure that agents remain shutdownable, we have to ensure that these agents maintain their preferential gaps. But it’s been argued that we should expect advanced artificial agents to complete their preferences (thereby resolving any preferential gaps), because having incomplete preferences makes these agents vulnerable to pursuing dominated strategies.
Sami and I have objected to these arguments. We claim that agents with incomplete preferences/preferential gaps can act in accordance with policies that keep them from pursuing dominated strategies. But there are still issues to address here. Here’s one. Agents acting in accordance with the relevant policies get trammelled to some extent: they act as if they’re resolving their preferential gaps over time. If these agents face a sufficiently wide variety of choices, they’ll eventually converge to behaving like agents with complete preferences. So a question: how much of a problem is this likely to be? See Sami’s post and my comment for some thinking here.
Here’s another issue. Goal-content integrity is a convergent instrumental subgoal of rational agents: agents have instrumental reason to maintain their strict preferences. But goal-content integrity seems to give agents no instrumental reason to maintain their preferential gaps.[31] So even though agents need not resolve their preferential gaps to avoid pursuing dominated strategies, there’s a risk that they resolve them anyway. And this risk seems to be exacerbated by the fact that resolving preferential gaps is one way to avoid pursuing dominated strategies, even though it’s not the only way.
But here’s a possible response that seems worth exploring further: Timestep Dominance itself will keep TD-agents from resolving their preferential gaps. Resolving preferential gaps will likely cost the agent (at least some small amount of) resources, and so will lead to lower expected sum-total utility conditional on shutdown at some timestep. And resolving preferential gaps won’t lead to greater expected sum-total utility conditional on shutdown at any timestep, because the only preferential gaps to resolve are between different-length trajectories, and gaining a preference between different-length trajectories won’t lead the agent to make choices that are better with respect to expected sum-total utility conditional on shutdown at a timestep. As a consequence, resolving preferential gaps will be timestep-dominated by not resolving preferential gaps. See here for some early thinking on this front.
21.2. Audere’s theorem
Theorems like Audere’s present a challenge to candidate solutions to the shutdown problem. Audere lays out two conditions on preferences that (seemingly) agents must satisfy in order to be shutdownable and useful. Audere then proves that any agent satisfying these conditions has cyclic preferences. There’s still work to do here but I think that this challenge is surmountable. See this document for more detail.[32]
Audere’s theorem also suggests a more general project: tracing the implications of principles like POST, Timestep Dominance, and possible additional principles. Perhaps it can be proved that some of these implications are undesirable, in which case we’ll need to consider what to do in light of them.
21.3 Managing the news
Many proposed solutions to the shutdown problem incentivise agents to ‘manage the news’: to ensure that the shutdown button is pressed conditional on bad news and not pressed conditional on good news (see section 4.2 of Soares et al.). I suggest a way to prevent TD-agents from managing the news in my contest entry (in section 15), but I suspect that there are better ways to be found.
Here’s one reason for optimism. On shutdown proposals like Stuart Armstrong’s utility indifference, agents are compelled to manage the news by the very feature that keeps them shutdownable. Armstrong’s proposal thus faces a dilemma: either the agent isn’t shutdownable, or the agent manages the news.
But if (as seems likely) managing the news will cost the agent at least some small amount of resources, the IPP avoids this dilemma. Timestep Dominance keeps the agent shutdownable, but Timestep Dominance doesn’t compel the agent to manage the news if doing so costs resources. So plausibly we can find additional principles that will prevent the agent from managing the news.
21.4. Incidental button-manipulation
I argued above that TD-agents won’t resist shutdown, because Not Resisting Always Timestep-Dominates Resisting. But that only rules out TD-agents trying to prevent the pressing of the shutdown button. It doesn’t rule out TD-agents preventing the pressing of the shutdown button incidentally, as a byproduct of doing other things. Here’s what I wrote when I first mentioned this issue, in footnote 22 above:
21.5. Multi-agent dynamics
I haven’t given multi-agent dynamics much thought yet, but there could be problems here. It’s possible that TD-agents are easily manipulable by other agents. For example, perhaps other agents have to pay only very small costs to get TD-agents to do undesirable things. If so, we’d need to consider the extent of the problem and what to do in light of it.
21.6. Training TD-agents
My proposed training regimen is speculative and imprecise in various ways. Could it work? If so, how should we make it more precise?
What are some quick and cheap ways to test the promise of the IPP? How can we build on this experiment?
21.7. Maintaining the shutdown button and creating corrigible subagents
The IPP aims to ensure that agents are shutdownable: that they shut down when we want them to shut down. This notion of shutdownability differs from Soares et al.’s notion of corrigibility. Per Soares and coauthors, corrigibility requires not only shutdownability but also that the agent fixes the shutdown button when it breaks, lets us modify its architecture, and only creates subagents that are themselves corrigible in the aforementioned ways.
I expect that this kind of corrigibility will require training agents to adopt ‘Fix the shutdown button when it breaks’ and ‘Only create corrigible subagents’ as terminal goals.[33] And there are three familiar problems that we confront in getting agents to adopt these as terminal goals: reward misspecification, goal misgeneralization, and deceptive alignment. And these three problems might be difficult to overcome in this case.
But I think that the IPP at least softens the latter two problems. To see why, consider first the terminal goal ‘Fix the shutdown button when it breaks’. And compare two kinds of agent. The first is an agent that prefers lotteries in line with their expectations of money made overall. Call this a ‘money-maximising agent’. The second is an agent that satisfies the two principles of the IPP – POST and Timestep Dominance – with utility at a timestep given by the amount of money that the agent makes at that timestep. Call this an ‘IPP-agent’.
Money-maximising agents have two incentives to avoid adopting ‘Fix the shutdown button when it breaks’ as a terminal goal:
IPP-agents, by contrast, only have the first incentive to avoid adopting ‘Fix the shutdown button when it breaks’ as a terminal goal. These agents don’t have the second incentive, because these agents don’t care about increasing the probability of early shutdown. That’s because these agents satisfy POST and so lack a preference between every pair of different-length trajectories.
Now suppose that we successfully inculcate into these agents some impulse towards fixing the shutdown button, but that satisfying this impulse isn’t perfectly correlated with actually fixing the button. Suppose, for example, that actually fixing the button requires connecting each of ten crucial wires but that the agent can fully satisfy its impulse by connecting just nine of them. Money-maximising agents have strong incentives to do this kind of fake-fixing rather than actually-fixing, because fake-fixing satisfies the agent’s impulse without increasing the probability that the agent gets shut down early.
By contrast (and supposing that fake-fixing and actually-fixing are equally resource-intensive), IPP-agents have no incentive to fake-fix rather than actually-fix. That’s because (as noted above) IPP-agents don’t care about increasing the probability of early shutdown. This fact suggests that the impulses towards button-fixing inculcated into IPP-agents are likely to generalise better than the impulses towards button-fixing inculcated into money-maximising agents. The IPP thus softens the problem of goal misgeneralization that we confront in trying to get agents to adopt ‘Repair the shutdown button when it breaks’ as a terminal goal.
Now consider deceptive alignment. As noted above, IPP-agents only have the first incentive to avoid adopting ‘Fix the shutdown button when it breaks’ as a terminal goal:
This first incentive seems like it could be relatively weak: fixing the shutdown button needn’t be that costly. And if the incentive is weak, then IPP-agents only have weak incentives to behave deceptively in training in an attempt to avoid being made to adopt ‘Repair the shutdown button when it breaks’ as a terminal goal. And then plausibly there’s some way that we can set up the training process so that the agent’s overall balance of incentives favours being honest about whether it has ‘Repair the shutdown button when it breaks’ as a terminal goal. For example, we can present the IPP-agent with “honest tests” (see Carlsmith) like seeing what the IPP-agent does when we offer it some chunk of compute to spend on anything it wants. So the IPP also seems to soften the problem of deceptive alignment that we confront in trying to get agents to adopt ‘Repair the shutdown button when it breaks’ as a terminal goal.
Now consider the terminal goal ‘Only create corrigible subagents’. Here the situation is slightly different. I think we’ll need to find some way to train IPP-agents not to care about what happens after they’ve been shut down. If we can achieve that, then IPP-agents have little reason to create incorrigible subagents rather than corrigible subagents. Here’s why. Incorrigible subagents only start behaving significantly differently to corrigible subagents after we humans start wanting the agent and its subagents to shut down. The significant difference in behaviour is: corrigible subagents would shut down at this point, whereas incorrigible subagents wouldn’t.
IPP-agents will understand that they themselves will likely be shut down soon after we humans start wanting them to shut down (because IPP-agents will understand that they themselves satisfy Timestep Dominance and so won’t resist shutdown). If these IPP-agents don’t care about what happens after they’ve been shut down, then they don’t care about the only period of time in which incorrigible subagents and corrigible subagents behave significantly differently.
IPP-agents thus have little reason to create incorrigible subagents rather than corrigible subagents. And if that’s right, then (similarly to my discussion of ‘Fix the shutdown button when it breaks’ above):
The IPP thus seems to soften the problems of goal misgeneralization and deceptive alignment that we confront in training agents to adopt ‘Create only corrigible subagents’ as a terminal goal.
Note that I’m still uncertain about much that I’ve written in this last subsection. That’s why it’s under the heading ‘Issues still to address’. I’d be interested to hear about possible problems and suggestions for other strategies.
22. Conclusion
The post is long, but the Incomplete Preferences Proposal (IPP) is simple. We keep artificial agents shutdownable by training them to satisfy two principles. The first is:
And the second is:
Where ‘timestep-dominates’ is roughly defined as follows:
POST paves the way for Timestep Dominance, and Timestep Dominance keeps agents shutdownable because Not Resisting Always Timestep-Dominates Resisting. Timestep Dominance also allows agents to be useful because it only rules out timestep-dominated lotteries. It leaves most preferences open, to be decided by some other principle.
POST and Timestep Dominance seem trainable too. We could train agents to satisfy these principles by making small changes to an otherwise-thoroughly-prosaic setup for training advanced AI. For POST, we give agents lower reward for repeatedly choosing same-length trajectories. For Timestep Dominance, we give agents lower reward for choosing timestep-dominated lotteries.
By training agents to satisfy the IPP in this way, we seem to largely circumvent the problems of reward misspecification, goal misgeneralization, and deceptive alignment. On reward misspecification, the IPP seems easy to reward. On goal misgeneralization, the IPP seems simple. On deceptive alignment, the IPP seems never to give agents a chance to learn goals that incentivise preventing shutdown in deployment.
For discussion and feedback, I thank Yonathan Arbel, Adam Bales, Ryan Carey, Eric Chen, Bill D’Alessandro, Sam Deverett, Daniel Filan, Tomi Francis, Vera Gahlen, Dan Gallagher, Jeremy Gillen, Riley Harris, Dan Hendrycks, Leyton Ho, Rubi Hudson, Cameron Domenico Kirk-Giannini, Jojo Lee, Jakob Lohmar, Andreas Mogensen, Murat Mungan, Sami Petersen, Arjun Pitchanathan, Rio Popper, Brad Saad, Nate Soares, Rhys Southan, Christian Tarsney, Teru Thomas, John Wentworth, Tim L. Williamson, Cecilia Wood, and Keith Wynroe. Thanks also to audiences at CAIS, GPI, Oxford MLAB, Hong Kong University, the AI Futures Fellowship, and EAG Bay Area.
Note that we need agents to be both shutdownable and useful. If the best we can do is create an agent that is only shutdownable, we still have to worry about some AI developer choosing to create an agent that is only useful.
See Pan, Bhatia, and Steinhardt. Related is the problem of outer alignment.
See Shah et al. and Langosco et al. Related is the problem of inner alignment.
See Hubinger et al. and Carlsmith.
Note that my notion of shutdownability differs slightly from Soares et al.’s (2015, p.2) notion of corrigibility. As they have it, corrigibility requires not only shutdownability but also that the agent repairs the shutdown button, lets us modify its architecture, and continues to do so as the agent creates new subagents and self-modifies. See section 21.7 for some thoughts on how we might get those extra features.
This definition differs from another way you might define ‘lack of preference’: an agent lacks a preference between lottery X and lottery Y iff they wouldn’t trade one for the other. They’d stick with whichever lottery they had already. See John Wentworth’s Why subagents? And John Wentworth’s and David Lorell’s Why not subagents?
See Joseph Raz and Ruth Chang for examples along these lines.
My definitions of ‘indifference’ and ‘preferential gaps’ differ slightly from the usual definitions. Each is usually defined in terms of weak preference as follows. An agent is indifferent between X and Y iff it weakly prefers X to Y and weakly prefers Y to X. An agent has a preferential gap between X and Y iff it doesn’t weakly prefer X to Y and doesn’t weakly prefer Y to X. I depart from the usual definitions because weak preference makes for a bad primitive in this context: it’s complicated to define in terms of dispositions to choose.
In any case, my notions of indifference and preferential gaps are closely related to the usual notions. If we assume that weak preference is transitive, the usual definition of indifference implies my definition, and my definition of preferential gaps implies the usual definition. If we add a domain-richness condition (if X is not weakly preferred to Y and Y is not weakly preferred to X, then there is some sweetening of X that is not weakly preferred to Y, or some sweetening of Y that is not weakly preferred to X, or some souring of X to which Y is not weakly preferred, or some souring of Y to which X is not weakly preferred), then the usual definition of indifference is biconditional with my definition, and the usual definition of preferential gaps is biconditional with my definition.
It’s a variation on the Second Theorem in my contest entry and in An AI Engineering Puzzle for Decision Theorists. John Wentworth makes a similar point in this post.
In An AI Engineering Puzzle for Decision Theorists, I define ‘indifference’ and ‘preferential gaps’ in terms of weak preference, and so the Second Theorem includes Transitivity as an antecedent condition. In this post, the requisite instances of Transitivity are built into my definitions of ‘indifference’ and ‘preferential gaps’ (see the previous footnote), so Transitivity isn’t an antecedent condition of this theorem.
But see Goldstein and Robinson for an argument that these agents could be useful.
See Soares et al. 2015, p.2; Ngo, Chan, and Mindermann 2023, p.7, and 'Nearest unblocked strategy'
The agent might also learn instead a false belief that button manipulation doesn’t work. We’d then have to worry about the agent discovering the truth in deployment. See ‘Beliefs and goals can be mixed together’ in Barnett and Gillen.
The agent might also learn instead a mere ‘outer shell’ constraint against manipulating the button, analogous to some humans’ instinctual fear of heights. The agent might regard this kind of constraint as external to its terminal goals: as simply an obstacle to be overcome. We’d then have to worry about the agent overcoming the constraint in deployment. See ‘Outer shell non-consequentialist constraints’ in Barnett and Gillen.
‘The shutdown button never gets pressed’ can be treated as just one more possible length.
See Towards Shutdownable Agents Via Stochastic Choice for experimental results suggesting that the idea is promising.
There is a respect in the which the accountant-clown analogy breaks down. Your lack of preference between the two careers is insensitive to some sweetenings and sourings, but it isn’t insensitive to all sweetenings and sourings. You’d prefer being a clown with salary $n+1,000,000,000 to being an accountant with salary $m. By contrast, POST requires lacks of preference that are insensitive to all sweetenings and sourings: no sweetening or souring can induce a preference between different-length trajectories.
I’m not worried about this disanalogy. Supposing that preferential gaps are possible at all, I see no reason to think that they have some maximum possible size. In any case, we could test this hypothesis using (something like) my proposed training method or the experiments in this paper.
That said, there remain issues to address here. For example, although agents with preferential gaps can make themselves invulnerable to pursuing dominated strategies by choosing in certain ways, they can also make themselves invulnerable by completing their preferences, so how do we ensure that they do the former rather than the latter? I discuss this and similar issues in section 21.1.
Here’s an alternative reason why you might doubt that preferential gaps are maintainable: you might think that all preferential gaps depend on the agent having some uncertainty about the nature of the objects of preference. You’d then think that the agent resolving its uncertainty would also lead the agent to resolve its preferential gaps and thereby complete its preferences.
I think that some preferential gaps depend on uncertainty, but not all. Consider an example. You’re offered a choice between two very different flavours of ice-cream. You might know everything that there is to know about how these two flavours will taste, what their nutritional content is, etc., and yet still lack a preference between the two flavours, and this lack of preference might be insensitive to some (literal, in this case) sweetening or souring.
So it seems possible for an agent to have preferential gaps that don’t depend on uncertainty. And (it seems to me) my proposed method for training agents to satisfy POST will bring about preferential gaps that don’t depend on uncertainty. That’s because these agents can be trained to satisfy POST in environments in which there’s no uncertainty about the trajectories that will result from the agent’s actions.
Or, more precisely: for each timestep, Leave maximises expected sum-total utility conditional on shutdown at that timestep. This phrase is cumbersome, so I’ll continue to use ‘conditional on shutdown at each timestep’ as an abbreviation for the more precise thing that I really mean.
Here’s an extra requirement for lottery X timestep-dominating lottery Y that seems implicit enough and complicating enough to leave out of the main text: for each timestep t, lottery X assigns positive probability to shutdown occurring at timestep t iff lottery Y also assigns positive probability to shutdown occurring at timestep t. In light of this extra requirement, conditions (1) and (2) should be modified to begin with ‘Conditional on shutdown at each [/some] timestep with positive probability of shutdown occurring at that timestep.’ This extra requirement prevents us conditioning on zero-probability events. The extra requirement will be satisfied for each pair of lotteries X and Y by any agent with a regular probability function, defined as a probability function that assigns probability 1 only to logical truths and probability 0 only to contradictions (see Hájek). Agents with regular probability functions will always assign positive probability to shutdown at each timestep.
The timestep dominance relation in the main text is a specific instance of a more general timestep dominance relation:
The specific timestep dominance relation in the main text follows from this general timestep dominance relation plus the following principle:
We could get other specific instances of the timestep dominance relation by substituting in other principles to govern the agent’s preferences over same-length sublotteries. For example, the agent could be risk-averse with respect to sum-total utility. This flexibility seems like an advantage of the proposal because some other principle might make Not Resisting Always Timestep-Dominates Resisting (explained below) more plausible. I stick with the ‘expected sum-total utility’ version in the main text for ease of understanding.
Holtman (2020, p.15) also makes this kind of point.
I’ve argued in this paragraph that TD-agents won’t try to prevent the pressing of the shutdown button. A parallel argument suggests that TD-agents also won’t try to cause the pressing of the shutdown button: doing so will cost the agent resources, and those resources can’t also be spent on pursuing utility at a timestep.
The caveat is that I’ve defined ‘resisting shutdown’ as trying to prevent the pressing of the shutdown button. So defined, ‘resisting shutdown’ doesn’t include cases in which the agent acts in accordance with Timestep Dominance and just happens to prevent the pressing of the shutdown button as a byproduct.
Here’s an example of such a case. Consider a TD-agent whose goal is to make money. This agent builds a lemonade stand in the most convenient place for selling lemonade, which just happens to be on top of the shutdown button. As a result, we humans are prevented from pressing the button.
Sami Petersen has written an interesting discussion of cases like this. There’s more work to be done here, but I’m optimistic that this issue can be overcome. Here’s why. Misaligned agents are hard to control (in large part) because they’re trying to undermine our efforts (see section 1.2.4 of Carlsmith for a similar sentiment). But TD-agents aren’t trying to prevent us shutting them down: if they prevent us, it’ll only be as a byproduct of doing other things. And plausibly we can find some way to set things up so that acting in accordance with Timestep Dominance won’t happen to prevent the pressing of the shutdown button as a byproduct.
For example, it follows from Maximality, which says that an option may be chosen iff it’s not dispreferred to any other available option.
Sami Petersen has a proof that a general version of Timestep Dominance follows from POST plus a ‘Comparability Class Dominance’ (CCD) principle. An outcome o’s comparability class is defined as the set of all outcomes preferred, dispreferred, or indifferent to that outcome (i.e. the set of all outcomes not related to o by a preferential gap), and CCD says roughly: if the agent weakly prefers lottery X to lottery Y conditional on each comparability class, and strictly prefers X to Y conditional on some comparability classs, then the agent prefers X to Y.
Here I’m cribbing from training setups described by Cotra, Ngo, Chan, and Mindermann, and DavidW.
Here’s a concern you might have around this point. Given current neural network architectures, agents’ stochastic choosing is (in some sense) only skin-deep: the neural network itself is deterministic; its role ends when it outputs a vector of final-layer activations; some function transforms this vector into a probability distribution; and actions are chosen by sampling from this probability distribution. The stochasticity only comes in at the very end of the process, and this might make you worry that stochastic choosing doesn’t really indicate a lack of preference.
In reply, note two points. The first will be a dissatisfying semantic note. In this post, I’ve defined ‘lack of preference’ such that stochastic choosing entails a lack of preference, irrespective of how that stochastic choosing is implemented.
But this semantic note leaves open the important question: does stochastic choosing between different-length trajectories indicate that the agent won’t pay costs to shift probability mass between different-length trajectories? We need that latter thing in order for the agent to be shutdownable. (And also: one could quite reasonably define ‘lack of preference’ in such a way that ‘the agent won’t pay costs to shift probability mass between X and Y’ is a necessary condition for the agent lacking a preference between X and Y.)
I give my answer to the important question in section 18. In brief, I think that the answer is yes: stochastic choosing between different-length trajectories indicates that the agent won’t pay costs to shift probability mass between different-length trajectories. That’s because this claim follows from a more general principle that we’ll train agents to satisfy, and we’ll train agents to satisfy that more general principle because it’s a prerequisite for minimally sensible action under uncertainty. For more, see section 18.
Here are some early thoughts. To get agents to satisfy POST, we need to use a policy-gradient method rather than a value-based method. That’s because (ignoring exploratory moves) the policies learned by value-based methods are deterministic. And we need to train the agent in a POMDP in which the agent’s observations aren’t Markovian state signals. In particular, we need the conditional probability distribution over future rewards to depend on the agent’s actions in previous episodes, and we need to ensure that the agent can’t observe/remember their actions in previous episodes. If these latter conditions aren’t satisfied, then some deterministic policy will be among the optimal policies, and we don’t want that.
Carlsmith makes this kind of point here:
And here:
Doing so will cost utility at a timestep, and it will only shift probability mass between shutdowns at different timesteps. See Not Resisting Always Timestep-Dominates Resisting.
See John Wentworth’s Level 1 and the Second Theorem for similar arguments.
See the heading ‘Project ideas’ in this Google Doc for more detail here.
I’m planning to turn this Google Doc into a proper post soon.
See sections 13 and 14 of my contest entry for why.